Speech recognition model training method and device and speech recognition method and device

Abstract

The invention provides a voice recognition model training method and device and a voice recognition method and device. The voice recognition model training method comprises the steps: acquiring a voice sample pair; obtaining a first sample vector of the first voice sample and a second sample vector of the second voice sample through a full connection layer of the voice recognition model; calculating a loss value of the speech recognition model according to the similarity between the first sample vector and the second sample vector and the sample pair label; and according to the loss value of the speech recognition model, training the speech recognition model until a training stop condition is met, and obtaining a trained speech recognition model. According to the method, the loss value of the speech recognition model based on the sample pair is introduced, and the loss value enables the speech recognition model to have the capability of accurately extracting the speech content information, so that when the to-be-recognized keywords are newly added, the speech recognition model does not need to be retrained, and only a small amount of standard speech of the keywords needs to be prepared.

Description

technical field [0001] This specification relates to the technical field of speech processing, in particular to a speech recognition model training method and device, and a speech recognition method and device. Background technique [0002] With the rapid development and wide application of computer technology and network technology, there are many scenarios that require speech recognition. For example, in a voice word recognition game, it is necessary to detect whether there is a word (target keyword) displayed on the screen based on the user's voice; or, the device can be controlled by voice, so the device needs to recognize the user's voice to obtain the target keywords, and then trigger corresponding operation instructions based on the target keywords to operate the device. [0003] In the prior art, it is generally based on a convolutional neural network (CNN) to predict which category of a speech speech belongs to a given keyword category, so as to determine whether t...

AI Technical Summary

Problems solved by technology

[0004] However, in the above speech recognition method, the output of the convolutional neural network uses the normalized exponential function softmax function to perform classification prediction. The dimension of the softmax function is the number of given keywords. When adding keywords, the dimension of softmax also To increase accordingly, the model needs to be retrained. Since the training model needs to collect a certain amount of voice data of the newly added keyword, the time and labor costs are high.

In addition, after dividing the speech into frames and predicting the probability that each speech frame belongs to each given keyword, it is necessary to use a sliding window larger than the speech frame window to smooth the probability of the speech frame to obtain that the speech frame belongs to each given keyword. Given the confidence of the keyword, the specific smoothing strategy method and the window used for smoothing need to be adjusted according to the actual situation, and the process is relatively cumbersome

Images